Troubleshooting stereo depth corner noise in point clouds

A stereo point cloud can look stable near the image center yet become scattered in one corner, even when the target stays at the same distance. That pattern does not identify a single fault: weak texture, reflections, stereo overlap, calibration, and processing can produce similar symptoms. This troubleshooting sequence helps you separate those causes using controlled captures, then decide whether filtering improves useful measurements or merely removes difficult points.

1. Confirm where the noise first appears

Start with the native depth or disparity image, not only the rendered point cloud. A visualization can exaggerate isolated points, while color alignment can introduce borders that do not exist in the original depth output.

Record the following for every capture:

• Native stereo image resolution and depth output resolution.
• Exposure, gain, illumination state, and any automatic adjustments.
• Rectification, alignment, confidence rejection, and filtering settings.
• Target position, orientation, and approximate distance.
• Whether the displayed cloud includes invalid or previously accumulated points.

Save synchronized left and right images when the documented interface permits it. Also save disparity, depth, and the validity mask if available. These reveal whether the error begins during correspondence matching or appears during later conversion and display.

If native depth looks clean but the cloud does not, inspect coordinate conversion and invalid-point handling first. The guide to checking missing XYZ points in stereo visualization addresses that separate stage.

2. Determine whether the problem follows the target or the image corner

Use a rigid, matte target with nonrepeating texture before testing a smooth cylinder. Keep its orientation approximately constant while moving it between the center and the affected corner.

Capture three positions: center, suspect corner, and the opposite corner. Repeat the sequence without changing settings. If practical, return to the center afterward to check that illumination or exposure has not drifted.

A triangular affected region is a useful clue, not proof of a particular defect. Compare its boundary against image padding, valid-overlap masks, reflections, and the target silhouette.

For curved parts, keep the fitted surface and the visible surface patch comparable. Moving a pipe sideways can change which part of its curved surface is visible, even at constant optical-axis depth.

3. Check texture and illumination before changing calibration

Stereo matching needs distinguishable features visible to both views. Smooth plastic, repetitive patterns, and glossy surfaces can make several candidate matches look equally plausible.

Inspect the suspect region in both source images at native resolution. Look for low contrast, saturation, motion blur, dark shading, and reflections that appear differently in each view. A scene that looks adequately lit to a person may still produce poor images at the selected exposure.

Try these tests one at a time:

1. Replace the production part with a matte, textured reference at similar depth.
2. Add diffuse illumination without pointing a bright source into either lens.
3. Adjust exposure to retain texture without clipping bright regions.
4. Lock exposure and gain after establishing a usable image, where supported.
5. If the system has documented active illumination controls, compare supported operating states without modifying the optical hardware.

Use removable surface treatment only on a sacrificial sample where the process permits it. It is a diagnostic comparison, not evidence that untreated production parts will work.

If brighter illumination makes the depth worse, investigate saturation and active-illumination interference rather than adding more light. The article on IR depth noise under bright illumination covers that different failure pattern.

4. Check stereo overlap and edge-of-field geometry

A point must be visible in both stereo views to support an ordinary correspondence measurement. The full image from either individual camera is not necessarily the usable depth field.

At close range, the common viewing region can be substantially smaller than either source image. Part edges can also be visible in one view but hidden in the other. A cylindrical target is especially useful for exposing this distinction because its silhouette creates changing occlusion boundaries.

The explanation of why stereo depth coverage is smaller than the source image helps distinguish expected coverage loss from excessive measurement noise.

Inspect whether the suspect points occur:

• Outside the documented valid depth region.
• Along a silhouette or depth discontinuity.
• Near rectification padding or a cropped image boundary.
• On strongly sloped surfaces with little usable texture.
• Near the minimum supported operating distance.

Do not assume every corner has equivalent performance. Nevertheless, a strong asymmetry on a flat textured target deserves investigation rather than automatic cropping.

The related guide to stereo depth errors near object edges explains why boundary behavior should be assessed separately from surface-interior precision.

5. Use the stereo error equation correctly

For a rectified stereo pair, an idealized depth relationship is:

Here, Z is optical-axis depth, f is focal length expressed in pixels, B is baseline, and d is disparity in pixels. Baseline and depth must use the same length unit. The second expression is a first-order estimate for small disparity uncertainty, not a model of gross correspondence failures.

Consider an illustrative system with a focal length of 600 pixels and a baseline of 60 mm. At a depth of 350 mm:

The example assumes disparity uncertainty of 0.25 pixels. At 700 mm, the same assumptions give approximately 3.4 mm of depth uncertainty.

This explains increasing uncertainty with distance. It does not, by itself, explain deterioration in one corner at constant optical-axis depth. That spatial pattern means the matching uncertainty, geometry, validity, or processing behavior is changing across the image.

Also distinguish optical-axis depth from straight-line range to the camera. Points on a plane perpendicular to the optical axis share the same depth but have different ranges away from the image center.

6. Verify rectification and calibration compatibility

Investigate calibration after checking image quality and valid overlap. Calibration errors can leave corresponding features on different rows or make residual errors vary across the field.

For a conventional horizontally rectified pair, corresponding features should lie close to the same image row. Inspect the center and suspect corner using a stationary, textured target. Compare the observed alignment against the manufacturer's documented verification criteria rather than inventing a universal pixel tolerance.

Check that:

• Calibration belongs to the actual stereo assembly.
• Image resolution, crop, and binning match the calibration pipeline.
• Left and right streams have not been swapped.
• Images are not rectified twice.
• Intrinsic parameters match any resizing or cropping operations.
• The stereo assembly has not shifted mechanically.

The raw stereo calibration and metadata checks provide a useful acquisition baseline. If processing recorded images, follow the separate workflow for rectifying stored images with their calibration data.

Use only the manufacturer's documented calibration or recovery procedure. Do not compensate for a suspected calibration fault by manually shifting one image until a single scene looks better.

7. Compare rejection and filtering without hiding the failure

Create a baseline with optional postprocessing disabled or minimized through supported settings. Preserve required rectification and the manufacturer's valid operating configuration.

Then change one control at a time. Confidence rejection, consistency checks, spatial smoothing, temporal filtering, and decimation solve different problems.

Confidence values do not have a universal direction. A higher setting may be stricter in one interface and more permissive in another. Read the parameter definition and confirm the effect on both point retention and outliers.

Use two measurements together:

• Valid-point fraction: valid samples divided by all samples in the evaluation region.
• Surface residual: distance from retained points to an appropriate reference surface, using the same fitting and evaluation method.

For example, suppose a hypothetical corner region contains 10,000 samples. A baseline retains 7,200 samples, giving 7,200 / 10,000 = 72% valid coverage. A stricter setting retains 4,800 samples, giving 48% coverage.

If the chosen residual statistic falls from 1.2 mm to 0.7 mm, report that improvement alongside the loss of 24 percentage points in coverage. A smoother cloud is not automatically a better inspection result.

Median-based decimation may suppress isolated outliers, but it also reduces spatial sampling. The guide to unexpectedly reduced depth resolution helps verify the output dimensions before comparing detail.

8. Verify the result against the actual inspection task

Use a flat reference for spatial comparisons and representative production parts for acceptance. A plane exposes corner-dependent distortion; a cylinder reveals whether the required curved surface remains measurable.

For each position, capture a short stationary sequence, such as 30 frames. Evaluate temporal variation separately from spatial surface error. A stable bias can pass a repeatability check while still producing an incorrect measurement.

Compare the center and corner using equal-sized regions that exclude silhouettes. Check depth bias, residual spread, valid coverage, and the actual feature measurement. Keep units and point-selection rules identical across configurations.

For cylinders, avoid judging quality only by a fitted radius. A fit can conceal missing surface sectors or systematic distortion. Report visible angular coverage and residual distribution as well.

If points remain missing rather than merely scattered, continue with the workflow for diagnosing holes in a point cloud.

Acceptance limits must come from the inspection requirement. A setting that supports coarse localization may be unsuitable for dimensional measurement, even when both point clouds look acceptable on screen.

9. Decide whether a different acquisition approach is justified

Do not replace hardware until the controlled tests identify a capability gap. Browse 3D sensors by acquisition method, then use the camera finder to narrow the available options.

The MV-DLS300P-04 is identified in the catalog as a laser galvanometer stereo camera. It is a candidate for application review, not a guaranteed remedy for corner noise.

The MV-DPS200P-02 is identified as a projected structured-light camera. Evaluate representative surfaces and the complete acquisition sequence before selecting it.

The supplied catalog descriptions do not establish their working-distance limits, corner accuracy, or suitability for a particular pipe surface. Request those details and an application assessment before ordering.

Building a stereo rig from separate area scan cameras is a different integration project. Matching optics, synchronization, mechanical stability, and calibration become system responsibilities; choosing cameras alone does not resolve them.

10. Close the investigation with a repeatable checklist

Before accepting a filter preset or requesting replacement hardware, confirm that you have:

• Reproduced the corner defect on a stationary target.
• Compared a textured reference with the production surface.
• Inspected both source images for clipping, blur, and weak texture.
• Separated native depth from aligned depth and cloud visualization.
• Checked valid stereo overlap and occlusion boundaries.
• Verified calibration compatibility with the active stream geometry.
• Compared minimally processed and filtered outputs.
• Reported point retention together with measurement error.
• Tested the final settings at every required target position.

Cropping is reasonable when the excluded region is outside the required inspection envelope. It is not an adequate fix when the part must be measured there. Likewise, filtering is appropriate when it meets the task's accuracy and coverage limits, not simply because the display looks cleaner.

For selection support, send representative source images, depth output, settings, and acceptance limits through application support. If the tests show that a different acquisition method is needed, start with the camera configuration workflow.